Re: [PATCH] scsi_dh_alua: handle RTPG sense code correctly during state transitions

[Hannes Reinecke <hare@suse.de>]

On 10/7/19 10:45 PM, Ewan D. Milne wrote:
> See below.
> 
> On Mon, 2019-10-07 at 15:57 +0200, Hannes Reinecke wrote:
>> From: Hannes Reinecke <hare@suse.com>
>>
>> Some arrays are not capable of returning RTPG data during state
>> transitioning, but rather return an 'LUN not accessible, asymmetric
>> access state transition' sense code. In these cases we
>> can set the state to 'transitioning' directly and don't need to
>> evaluate the RTPG data (which we won't have anyway).
>>
>> Signed-off-by: Hannes Reinecke <hare@suse.com>
>> ---
>>  drivers/scsi/device_handler/scsi_dh_alua.c | 21 ++++++++++++++++-----
>>  1 file changed, 16 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c b/drivers/scsi/device_handler/scsi_dh_alua.c
>> index 4971104b1817..f32da0ca529e 100644
>> --- a/drivers/scsi/device_handler/scsi_dh_alua.c
>> +++ b/drivers/scsi/device_handler/scsi_dh_alua.c
>> @@ -512,6 +512,7 @@ static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg)
>>  	unsigned int tpg_desc_tbl_off;
>>  	unsigned char orig_transition_tmo;
>>  	unsigned long flags;
>> +	bool transitioning_sense = false;
>>  
>>  	if (!pg->expiry) {
>>  		unsigned long transition_tmo = ALUA_FAILOVER_TIMEOUT * HZ;
>> @@ -572,13 +573,19 @@ static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg)
>>  			goto retry;
>>  		}
>>  		/*
>> -		 * Retry on ALUA state transition or if any
>> -		 * UNIT ATTENTION occurred.
>> +		 * If the array returns with 'ALUA state transition'
>> +		 * sense code here it cannot return RTPG data during
>> +		 * transition. So set the state to 'transitioning' directly.
>>  		 */
>>  		if (sense_hdr.sense_key == NOT_READY &&
>> -		    sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)
>> -			err = SCSI_DH_RETRY;
>> -		else if (sense_hdr.sense_key == UNIT_ATTENTION)
>> +		    sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a) {
>> +			transitioning_sense = true;
>> +			goto skip_rtpg;
>> +		}
>> +		/*
>> +		 * Retry on any other UNIT ATTENTION occurred.
>> +		 */
>> +		if (sense_hdr.sense_key == UNIT_ATTENTION)
>>  			err = SCSI_DH_RETRY;
>>  		if (err == SCSI_DH_RETRY &&
>>  		    pg->expiry != 0 && time_before(jiffies, pg->expiry)) {
>> @@ -666,7 +673,11 @@ static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg)
>>  		off = 8 + (desc[7] * 4);
>>  	}
>>  
>> + skip_rtpg:
>>  	spin_lock_irqsave(&pg->lock, flags);
>> +	if (transitioning_sense)
>> +		pg->state = SCSI_ACCESS_STATE_TRANSITIONING;
>> +
>>  	sdev_printk(KERN_INFO, sdev,
>>  		    "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",
>>  		    ALUA_DH_NAME, pg->group_id, print_alua_state(pg->state),
> 
> The patch itself looks OK, but I was wondering about a couple of things:
> 
>   - There are other places in scsi_dh_alua where the ASC/ASCQ 04 0A is checked
>     and we retry, I understand that this is a particular case you are solving
>     but is the changing of the state to -> transitioning (because that's what
>     the device said the state was) applicable in those other cases?
No. The original code was built around the assumption that RTPG would
return the status of the device; consequently we would have to retry
RTPG until we get a final status. But as mentioned, there are arrays
which cannot return RTPG data during transitioning, so the code would
never be able to detect a transitioning state.
With this patch we set the state directly once the said sense code is
received.
But this applies _only_ to the RTPG command, as this is required to move
the state machine along.
None of the other commands are affected.

>   - The code originally seems to have been under the assumption that the
>     transitioning state was a transient event, so the retry would pick up
>     the eventual state.  Now, some storage arrays spend a long time in the
>     transitioning state, but if we don't send another command are we going to
>     get the sense (or the UA) that triggers entry to the eventual ALUA state?
> 
Note, there are two types of retries.
The one is the 'normal' command retry, where we resend a command a given
number of times to retrieve the final status.
This is precisely the error which caused this patch.

And then there is a scheduled retry; here we essentially poll the array
with sending RTPG in regular intervals until the 'transitioning' state
is gone. (Check for 'alua_rtpg()' and the handling of the SCSI_DH_RETRY
return value). With the patch we continue to trigger that second type of
retries, which will eventually clear the transitioning state.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      Teamlead Storage & Networking
hare@suse.de			                  +49 911 74053 688
SUSE Software Solutions Germany GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 247165 (AG München), GF: Felix Imendörffer